Scholarlink

<p>This study develops a Composite Eutrophication Index (CEI) based on four water-quality parameters to predict river eutrophication risk, using ten quantitative and qualitative water variables. The CEI index, which is formed by three chemicals—(i) nitrate plus nitrite, (ii) orthophosphate, (iii) total phosphorus, and one physical (iv) dissolved oxygen—parameters, is modeled with supervised bagging and boosting ensemble (Random Forest, XGBoost, CatBoost) and self-supervised contrastive (SimCLR, MoCo, SimSiam) machine learning methods using stratified splitting technique. Compared to the statistical MLR model using the TOPSIS method based on seven deviance, tendency, and similarity statistical metrics, the ensemble models provide more accurate predictions. CatBoost outperforms others (TOPSIS rank 1, RMSE = 0.059, <i>R</i><sup>2</sup> = 0.881), while contrastive models (e.g., SimCLR, TOPSIS rank = 4, RMSE = 0.078) and MLR (TOPSIS rank = 7, RMSE = 0.098) show limitations. By the aid of explainable artificial intelligence techniques, SHAP and PDP/ICE analyses reveal physiochemical parameters (pH and Total Kjeldahl Nitrogen) and hydrological factors (e.g., discharge, suspended sediment) as key drivers, underscoring the critical role of eutrophication in exacerbating harmful algal blooms and persistent water pollution.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating ensemble and contrastive learning with explainable AI for modeling eutrophication-driven algal risks

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Sultan K. Salamah,
  • Marwan Kheimi,
  • Mohammad Zounemat-Kermani

摘要

This study develops a Composite Eutrophication Index (CEI) based on four water-quality parameters to predict river eutrophication risk, using ten quantitative and qualitative water variables. The CEI index, which is formed by three chemicals—(i) nitrate plus nitrite, (ii) orthophosphate, (iii) total phosphorus, and one physical (iv) dissolved oxygen—parameters, is modeled with supervised bagging and boosting ensemble (Random Forest, XGBoost, CatBoost) and self-supervised contrastive (SimCLR, MoCo, SimSiam) machine learning methods using stratified splitting technique. Compared to the statistical MLR model using the TOPSIS method based on seven deviance, tendency, and similarity statistical metrics, the ensemble models provide more accurate predictions. CatBoost outperforms others (TOPSIS rank 1, RMSE = 0.059, R2 = 0.881), while contrastive models (e.g., SimCLR, TOPSIS rank = 4, RMSE = 0.078) and MLR (TOPSIS rank = 7, RMSE = 0.098) show limitations. By the aid of explainable artificial intelligence techniques, SHAP and PDP/ICE analyses reveal physiochemical parameters (pH and Total Kjeldahl Nitrogen) and hydrological factors (e.g., discharge, suspended sediment) as key drivers, underscoring the critical role of eutrophication in exacerbating harmful algal blooms and persistent water pollution.